Is it possible to get an overall measure of variance explained by different levels of a categorical predictor, without using a reference level?

The difficulty here is that a level of a categrical variable is not a variable --- it is an outcome of the variable. So this is a bit like asking if it is possible to get a measure of variance explained for a continuous variable for the outcome $X=3$.

While it is an odd question, I suppose that the answer is yes, it is possible (though it would no longer be the same model). What you would do is to remove the initial categorical/continuous variable and replace it with a single indicator variable for the outcome of interest. You can then use ANOVA to measure the variance explained by this indicator variable. This would still compare the outcome to a reference level, but the reference level would be all cases where that outcome does not occur. Of course, this is not the same model as you are using when you include the full categorical variable (which I'm assuming has ore than two outcomes), but that would be the appropriate comparison.

Here's one idea, which may very well match Ben's answer. First, obtain the "eff" contrasts, which comprise comparisons between each mean and the average of all the means:

> mod <- lm(Petal.Width ~ Species, data = iris) > library(emmeans) > EMM <- emmeans(mod, "Species") > (CON = contrast(EMM, "eff")) contrast estimate SE df t.ratio p.value setosa effect -0.953 0.0236 147 -40.343 <.0001 versicolor effect 0.127 0.0236 147 5.360 <.0001 virginica effect 0.827 0.0236 147 34.982 <.0001 P value adjustment: fdr method for 3 tests 

Now, there is a relation between $R^2$ and $F$: $$ R^2 = 1 - (1 + df_R\cdot F / df_E)^{-1} $$ where $df_R$ is the d.f. for regression and $df_E$ is the d.f. for error. In this case, we want to develop an $R^2$-like statistic for each level of Species. And note that each $t^2$ is an $F$ statistic with one numerator d.f. accordingly, calculate:

> F <- test(CON)$t.ratio^2 > 1 - 1 / (1 + F/147) [1] 0.9171609 0.1634979 0.8927607 

I'll claim these are like $R^2$ statistics. They sum to more than 1 because they are interdependent.